Study on fast improving KLDA criterion for MSTAR SAR feature extraction and recognition

Kernel linear discriminant analysis (KLDA) is essentially a nonlinear feature extraction criterion. This work uses KLDA to extract the feature of MSTAR SAR (moving and stationary target acquisition and recognition synthetic aperture radar) images, through which the recognition rate is high and the intrinsic azimuth sensitivity in synthetic aperture radar image is overcome. At the same time, the KLDA computation cost is too high on the condition of more training samples. In order to deal with it, a fast feature vector selection (FFVS) scheme is adopted that divides the total images into several groups according to the dissimilarity of target classes and poses in image. The FFVS can fast select a part of images from each group as a subset whose mapping in high dimension feature space forms a basis. Each sample and the projection operator can be expressed by a linear combination of the basis, so the size of KLDA kernel matrix is decreased and the computation cost is reduced. Experimental results show that a good recognition performance is achieved by using the hybrid algorithm that combines FFVS and KLDA.